Introduction
From science to art and every discipline in between have felt the impact of Artificial Intelligence. Some cheerfully welcome the changing landscape; yet some get blindsided and left struggling on the track.
However, Pandora’s Box, once open, cannot be closed again; whether you like it or not, Artificial Intelligence is here to stay, and thus we need to face it.
This is the background which the event commenced; Indigenous Wikipedia and Artificial Intelligence Workshop is a special occurrence initiated by Wikimedia Foundation and Ministry of Education, co-headed by Wikidata Taiwan Community and Se-cyun Wikipidiya Seediq to lead the discussion on the contemporary challenges and explore the opportunities and concerns with the community. One of which is that, Large Language Models (LLMs) which consist of basically all the AI products on the market today, can very easily generate English and Chinese content at lightning speed, and more terrifying is that they are improving at an even higher speed. At the current state, most of the free-tier services already possess the capability to generate a full, syntactically correct Wikipedia article on their own.
From Community
Wiki as a movement focused on knowledge sharing is amongst the first line of the shore to witness the full impact. Since 2024, basically every event, no matter the original topic, the discourse will inevitably be visited by the worries of Artificial Intelligence, which has the full potential to strike down everything we’ve built.
The main issue is 2 folded, one more dangerous than the other.
Hallucination
First of all is the beaten path of LLM hallucination. Everyone by now should be familiar with this innate incapability of LLMs. For people who haven’t been introduced to LLMs, hallucination is a common error in which the LLM ‘hallucinated’ pure fictional information on its own and runs with it. Without diving too deep into the technicality, LLMs, despite all the marketing wrappers, are next-word-guessing machines no different than the technology we’ve been using on our computers and smartphones. Its underlying mechanism is purely distributional and by chance. Therefore, hallucination and mistakes are built into the core of every LLM; all we can do is minimize it through constraint after constraint to pull the wild distribution tighter, but no matter what we do, at the current stage, hallucination is not a bug but a built-in feature.
Context Pollution
The second issue is a branch of hallucination but unique enough to earn its own paragraph; Context Pollution is when an LLM, while trying to generate the answer, pools from the wrong (or hallucinated) context, which leads to the answer becoming more and more inaccurate over time, and worse, it is a self-reinforcing loop.
During the workshop with the Seediq community, we identified an unnerving pattern of such examples.
A quick context dump: Seediqs is one of the officially recognized Indigenous People living in the middle mountain part of Taiwan. Currently speaking, there are 4 distinct dialect communities (among them Truku, Toda, and Tgdaya are recognized officially; Tuda still earning its position), however, all sharing the same language label.
During a live demo of what is a LLM and how they work, we asked the “Grok” (LLM developed by ‘X ‘ (originally Twitter)) to reference Seediq’s Wikipedia and translate a paragraph of Traditional Chinese into Seediq. Astonishingly, it successfully produced a somewhat coherent result. From the review of community members, the translation turns out somewhat acceptable, there are a few anomalies of irregular grammar and wording but overall understandable.
However, upon closer inspection, we spot a very weird behavior in which the LLM mashed up the features from different dialects into the paragraph, an obvious example of context-pollution. Without extended experiments, we cannot be sure whether it is an issue with unspecified prompt or the it is the RAG system cannot distinguish the dialects. But what we identified here is another reason, besides hallucination, why LLMs are a dangerous tool for knowledge disciplines; Without proper knowledge in the subject, there’s no chance for people to identify what is true or false.
I believe we don’t need much explanation to understand why this is extremely troublesome for everyone in information disciplines. Untrackable ‘lies’ quietly spreading through the distribution of knowledge channels is a disaster waiting to explode.
The Response
The voice amongst the communities is varied; however, we all recognize the potential dangerous LLMs pose to the knowledge sector. Especially taking the outstanding scalability of speed and volume into consideration, there’s a reasonable concern that, if unregulated, LLMs can overwhelm the whole system with content covered in landmines at every corner.
However, we also recognized the benefit of how much LLMs, as a technology, have opened up so many opportunities for communities to take on challenges that are far beyond their normal conditions. The empowerment we felt is very real.
It is the perfect double-edged situation we are facing here; there is no easy answer.
Finally, at the end of the talk with the Seediqs community, we reached a minimal consensus that, due to limited manpower to spare, Seediqs Wikipedia refuses AI-generated content at the current stage. The fear is simple: if AI-generated content is allowed, it will simply overwhelm the community with ambiguity and thus lead to a collapse of trust.
Therefore, before better means of regulation emerge, the policy stays at the simplest refusal of AI content.
We believe this is an important step toward responsible Artificial Intelligence usage. Instead of hiding our heads in the sand, we need to face the challenges, take it seriously, and act, even at the bare minimum of simple refusal. It is enormous, as it laid down the foundation of communication, grounded it in context, and built transparency so people have a line to follow when they choose to; furthermore, it allows a reference point for regulators to identify those with malicious intent.




